Date of Award
2026-05-01
Degree Name
Master of Science
Department
Computer Science
Advisor(s)
Martine Ceberio
Abstract
Neural networks produce predictions but not reliable measures of how much to trust them. A clinician handed an output such as "97% chance malignant" has no way to tell whether that 97% reflects the model's actual reliability on inputs like this one, or whether the same number would have appeared on an image the model has no business classifying. The standard approaches to uncertainty quantification - Bayesian neural networks, deep ensembles, and Monte Carlo dropout - are useful. However, sometimes the uncertainty we need to handle only comes as lower and upper bounds. This raises a natural question: what if we could use intervals? If the input lies within a known range, interval arithmetic produces an output range that contains every value the network could have produced over that input. However, the simplest approach - naive interval propagation - produces bounds that are valid by construction but loose. The question we ask is whether the width of those loose bounds, after training the network with interval-aware losses, carries an uncertainty signal that lines up with prediction correctness - that is, whether it is calibrated, in the machine-learning sense. We construct an interval-aware training pipeline in four stages, from point training with interval inputs through training under four interval-aware loss functions: midpointwidth MSE, ratio, deviation, and midpoint cross-entropy. We introduce three notions of calibration: midpoint expected calibration error, width-accuracy correlation, and coverage. A controlled 4 x 4 sweep over losses and output activations on binary classification on MNIST evaluates how these design choices interact. No single configuration dominates all three calibration notions. Some configurations train to high accuracy with weakly informative intervals; one loss collapses every interval to width zero; coverage is meaningful only under linear output. The framework - the pipeline, the four losses, and the calibration vocabulary - is offered as groundwork for future work on calibration mechanics and on methods for tighter intervals.
Language
en
Provenance
Received from ProQuest
Copyright Date
2026-05
File Size
81 p.
File Format
application/pdf
Rights Holder
Yahriel Isaac Guel
Recommended Citation
Guel, Yahriel Isaac, "Uncertainty In Neural Networks: An Interval Approach" (2026). Open Access Theses & Dissertations. 4690.
https://scholarworks.utep.edu/open_etd/4690